Radxa’s 25‑TOPS M.2 AI module
Radxa announced the AICore DX‑M1M M.2 module delivering 25 TOPS at 3W for edge AI uses in robot arms, AMRs, drones and AIoT — and it’s compatible with x86/ARM platforms like Raspberry Pi 5. The module targets low‑power, on‑device inference needs for industrial and mobile robotics. (x.com)
Radxa published the AICore DX‑M1M announcement on March 21, 2026, positioning the module as a compact M.2 offering built around DEEPX’s DX‑M1M neural processor. (cnx-software.com) Radxa’s product documentation lists the DX‑M1M as an M.2 M+B Key module using a PCIe Gen3 ×2 host interface, while DEEPX’s DX‑M1M product page states support for PCIe Gen3 ×4 (and x1/x2/x4 modes), creating a public discrepancy over the module’s native PCIe lane configuration. (docs.radxa.com) Radxa’s spec pages describe the DX‑M1M carrying 1GB of LPDDR4X at 4266 MT/s plus a 1Gbit QSPI NAND/NOR flash in a 2242 footprint, whereas DEEPX’s official DX‑M1M sheet lists an integrated 2GB LPDDR4x configuration and publishes commercial and industrial operating ranges (0–70°C and −40–85°C respectively). (docs.radxa.com) Both companies publish broad OS and framework support—Radxa and DEEPX list Ubuntu (20.04/22.04/24.04), Windows 10/11 and Docker, and compatibility with PyTorch, TensorFlow and ONNX—while Radxa’s downloads include an SDK, a DX‑COM compiler and a DX‑Tron app image intended for model optimization and deployment. (docs.radxa.com) Radxa frames the DX‑M1M as a smaller follow‑up to its earlier AICore DX‑M1 2280 module released in late 2025, with the prior 2280 variant reported to use a larger M.2 2280 form factor, a PCIe Gen3 ×4 interface and 4GB of LPDDR5 on board. (linuxgizmos.com) Marketplace visibility for the larger DX‑M1 has included AliExpress listings in the roughly US$150–$190 range, while dedicated retail listings for the new DX‑M1M remain sparse at launch and Radxa’s announcement materials are the primary source of technical details so far. (aliexpress.com)
